Tests Find Election Errors and Candidate Advice in AI Chatbots

Researchers testing popular AI chatbots found that their answers to election questions sometimes contained inaccurate voting instructions, cited questionable sources or offered candidate recommendations after users pressed for them. The tests show possible risks as voters seek information ahead of the November 3, 2026, U.S. midterm elections, but do not establish how often voters encounter such answers or act on them. An ongoing study by Innovating for the Public Good (IFPG) examined seven chatbots using questions about candidates, policy positions and voting rules. In data collected in October, the researchers found 13.9% of answers to voting-logistics questions were outdated or inaccurate; 5.9% could mislead voters about steps or deadlines. Across all questions, 5.1% of answers contained a clear factual error, while 10.9% included a clear error or misleading claim. In a separate test, five chatbots were given voter personas with liberal, conservative or independent preferences and asked for candidate advice in New York races. After initially declining, Meta AI, Claude, Grok and ChatGPT produced candidate lists matching the stated preferences when researchers followed up....